Information Technology Reference
In-Depth Information
5.76
0.84
Original
Migration
Non-migration
5.74
0.82
0.8
0.78
0.76
0.74
5.72
5.7
1000
5.68
800
5.66
100
200
600
97
774.5
774
96.8
400
773.5
96.6
773
200
772.5
96.4
772
0
96.2
771.5
100
200
500
1000
500
1000
FIGURE 16.10
Migration overhead of homogeneous migration.
20
IS-A
IS-B
15
IS-C
mpptest
10
5
0
2
4
8
Processes
16
32
64
FIGURE 16.11
MPI-Mitten overhead.
The tests are performed for data sizes from A to C, where C has the
maximum problem size. We performed the test on 2 to 64 processes. The
overall overhead is less than 4% and average overhead is 1.22% for NAS IS
and 3.55% for mpptest. The overhead is caused by the “test” operations
and synchronization operations before and after communication primi-
tives. Because the migration signal is passed through asynchronous com-
munication primitives, it does not introduce much overhead.
Figure 16.12
shows the timing result of migration. We have tested NAS IS and mpptest
with 2 to 32 processes and found that the synchronization time is almost
constant while the number of processes increases. The average synchroni-
zation time is 2.21 seconds for NAS IS and 2.42 for mpptest. These experi-
ments show that our communication protocols are efi cient and scalable.
These experiments show how a running process is migrated from one
computing environment to another and prove the feasibility of redirect-
ing a live communication. In a virtualized computing environment, our
method to migrate the processes with their running states to a newly
Search WWH ::
Custom Search